Which Tasks Should You Give AI? A Simple Sorting Test

Most advice about AI productivity is about how to use AI — better prompts, better tools, better workflows. Almost nobody talks about the decision that comes first and matters more: which tasks should you hand over at all?

Get this wrong and AI actively costs you time. You spend fifteen minutes prompting, twenty minutes fixing the output, and end up with something worse than if you had just written it yourself. Get it right and the same tool clears hours off your week. This post is the sorting test I use, plus the category of work AI quietly makes worse that nobody warns you about.

The 3-Question Test

Before giving any task to AI, ask these in order. The first "no" is your answer.

1. Can I check the output quickly? If verifying the answer takes as long as doing the task, delegation saves nothing. You have moved the work, not removed it.
2. Does being wrong here cost me little? A rough draft nobody sees is low-stakes. A message to a client, a medical decision, or a number in a report is not.
3. Is the value in the output, not the doing? Some tasks exist to produce a thing. Others exist to change you — learning, thinking, deciding. Delegating the second kind defeats the purpose.

Three yeses means hand it over confidently. One no means either keep it, or reshape the task until it passes.

The Sorting Table

Here is where common tasks actually land:

Give to AI	Keep for yourself
First drafts you will heavily edit	Final versions people will judge you by
Summarising long text you also have	Deciding what the summary means for you
Reformatting, renaming, converting	Choosing the format in the first place
Generating options to react to	Making the actual choice
Boring repetitive text (alt text, tags)	Anything requiring your reputation
Explaining a concept you will verify	Learning something you must remember
Rubber-duck debugging your thinking	The thinking itself

Notice the pattern in the right column: judgement, taste, memory, and accountability. Those are the things that do not transfer. Everything else is mechanical, and mechanical work is exactly what you should be offloading.

The Trap: Tasks AI Makes Worse

This is the part missing from most productivity guides. Some tasks feel perfect for AI but leave you worse off:

• Learning something you actually need to know. Having AI explain a concept is useful; having it do your practice problems is not. The struggle is the learning. If you will need this skill again next week, do it yourself the first few times.
• Deciding what matters. AI will happily organise your to-do list into neat categories. It cannot tell you which item will change your month. Priority-setting looks like admin but is actually judgement.
• Writing that carries your voice. A personal message, an apology, a post where your perspective is the value. Readers can feel the difference, and generic output damages exactly the thing that made it worth writing.
• Anything you cannot verify. Confident wrong answers are the biggest AI risk. If you have no way to check a claim, you have not saved time — you have added an unknown risk.
• Broken processes. Automating a messy workflow gives you a faster mess. Fix the process first, then speed it up.

That last one is worth sitting with. Speed applied to the wrong thing just gets you to the wrong place sooner.

How to Reshape a Task So It Passes

A "no" is not always the end. Often you can restructure the task until delegation works:

• Hard to verify? Narrow the scope. Instead of "research this topic," give AI your own sources and ask it to summarise only those — then every claim is checkable. Tools that cite your uploaded documents make this far safer, which is why the NotebookLM study workflow works so well for research.
• High stakes? Move AI one step earlier. Do not let it write the client email; let it list what the email should cover, then write it yourself.
• Needs your voice? Use AI as a critic, not an author. "Tell me what is unclear in this paragraph" keeps your words and adds an outside eye.
• Too vague to delegate? That usually means you have not decided what you want. Clarify it for yourself first — which is also what makes prompts work, as I covered in how to write AI prompts that actually work.

A Realistic Weekly Setup

You do not need a stack of tools. A workable setup for one person:

1. One general assistant for drafts, explanations, and options.
2. One document tool for research grounded in sources you provide.
3. One automation for a repetitive task you do every week — a script, a scheduled job, a template.

Then protect two blocks of time where AI is switched off entirely: one for thinking about direction, one for the work only you can do. The point of clearing mechanical work is to create room for that, not to produce more output for its own sake.

Practical Takeaway

List the five tasks that ate the most time last week. Run each through the three questions. You will usually find one obvious candidate to delegate and one you were wrongly handing over. Fix both this week. The people who get the most from AI are not the ones using it for everything — they are the ones who decided carefully what to keep.

FAQ

Which tasks should I not use AI for?

Anything where you cannot verify the output, where being wrong is costly, where the doing is the point (learning, deciding), or where your personal voice is the value. Also avoid automating a process that is already broken.

Does AI actually save time?

Only for tasks that pass the verification test. If checking the output takes as long as doing the work yourself, you have relocated effort rather than removed it. Time savings are real for drafts, summaries, reformatting, and repetitive text.

Should I use AI for learning or studying?

Use it to explain, quiz, and clarify — not to complete work you need the skill from. The effort of struggling with a problem is what creates memory. Let AI shorten confusion, not replace practice.

How many AI tools do I need?

Usually two or three. One general assistant, one tool that works with your own documents, and one automation for a weekly repetitive task. More tools mostly add switching cost, not capability.

How do I stop AI writing from sounding generic?

Do not ask it for the final version. Use it for structure, options, or critique, then write the actual sentences yourself. Generic output is usually a sign the task needed your voice and should not have been fully delegated.